Towards Affective Natural Language Generation: Empirical Investigations

This paper reports on attempts to measure the differing effects on readers’ emotions of positively and negatively “slanted” texts with the same basic message. The methods of “slanting” the texts are methods that could be used automatically by a Natural Language Generation (NLG) system. A pilot study and a main experiment are described which use emotion self-reporting methods from Psychology. Although the main experiment was formulated with the benefit of knowledge obtained from the pilot experiment and a text validation study, nevertheless it was unable to show clear, statistically significant differences between the effects of the different texts. We discuss a number of possible reasons for this, including the possible lack of involvement of the participants, biases in the self-reporting and deficiencies of self-reporting as a way of measuring subtle emotional effects.